Rademacher functions

Jordan Bell
July 16, 2014

1 Binary expansions

Define S:{0,1}[0,1] by

S(σ)=k=1σk2k,σ{0,1}.

For example, for σ1=0 and σ2=1,σ3=1,,

S(σ)=02+14+18+=12;

for σ1=1 and σ2=0, σ3=0,,

S(σ)=12+04+08+=12.

Let σ{0,1}. If there is some n such that σn=0 and σk=1 for all kn+1, then defining

τk={σkkn-11k=n0kn,

we have

S(σ)=k=1n-1σk2k+k=n+112k=k=1n-1σk2k+12n=S(τ).

One proves that if either (i) there is some n such that σn=0 and σk=1 for all kn+1 or (ii) there is some n such that σn=1 and σk=0 for all kn+1, then S-1(S(σ)) contains exactly two elements, and that otherwise S-1(S(σ)) contains exactly one element.

In words, except for the sequence whose terms are only 0 or the sequence whose terms are only 1, S-1(S(σ)) contains exactly two elements when σ is eventually 0 or eventually 1, and S-1(S(σ)) contains exactly one element otherwise.

We define ϵ:[0,1]{0,1} by taking ϵ(t) to be the unique element of S-1(t) if S-1(t) contains exactly one element, and to be the element of S-1(t) that is eventually 0 if S-1(t) contains exactly two elements. For k we define ϵk:[0,1]{0,1} by

ϵk(t)=ϵ(t)k,t[0,1].

Then, for all t[0,1],

t=S(ϵ(t))=k=1ϵk(t)2k, (1)

which we call the binary expansion of t.

2 Rademacher functions

For k, the kth Rademacher function rk:[0,1]{-1,+1} is defined by

rk(t)=1-2ϵk(t),t[0,1].

We can rewrite the binary expansion of t[0,1] in (1) as

k=1rk(t)2k=k=1(12k-2ϵk(t)2k)=1-2k=1ϵk(t)2k=1-2t. (2)

Define r:{-1,+1} by

r(x)=(-1)[x],

where [x] denotes the greatest integer x. Thus, for 0x<1 we have r(x)=1, for 1x<2 we have r(x)=-1, and r has period 2.

Lemma 1.

For any nN,

rn(t)=(-1)[2nt]=r(2nt),t[0,1]

In the following theorem we use the Rademacher functions to prove an identity for trigonometric functions.11 1 Mark Kac, Statistical Independence in Probability, Analysis and Number Theory, p. 4, §3.

Theorem 2.

For any nonzero real x,

k=1cosx2k=sinxx.
Proof.

Let n and let c1,,cn. The function

k=1nckrk

is constant on each of the intervals

[s2n,s+12n),0s2n-1. (3)

There is a bijection between Δn={-1,+1}n and the collection of intervals (3). Without explicitly describing this bijection, we have

01exp(ik=1nckrk(t))𝑑t = s=02n-1s2-n(s+1)2-nexp(ik=1nckrk(t))𝑑t
= δΔn12nexp(ik=1nδkck)
= δΔnk=1neiδkck2
= k=1neick+e-ick2,

giving

01exp(ik=1nckrk(t))𝑑t=k=1ncosck. (4)

We have

01eix(1-2t)𝑑t=eixe-2ixt-2ix|01=eix(e-2ix-2ix+12ix)=sinxx. (5)

Using (2) we check that the sequence of functions k=1nrk(t)2k converges uniformly on [0,1] to 1-2t, and hence using (5) we get

01exp(ixk=1nrk(t)2k)dt01eix(1-2t)dt=sinxx

as n. Combining this with (4), which we apply with ck=x2k, we get

k=1ncosx2ksinxx

as n, proving the claim. ∎

We now give an explicit formula for the measure of those t for which exactly l of r1(t),,rn(t) are equal to 1.22 2 Mark Kac, Statistical Independence in Probability, Analysis and Number Theory, pp. 8–9. We denote by μ Lebesgue measure on . We can interpret the following formula as stating the probability that out of n tosses of a coin, exactly l of the outcomes are heads.

Theorem 3.

For nN and 0ln,

μ{t[0,1]:r1(t)++rn(t)=2l-n}=12n(nl).
Proof.

Define ϕ:[0,1] by

ϕ(t)=12π02πeix(-(2l-n)+k=1nrk(t))𝑑x

But for m,

12π02πeimx𝑑x=δm,0={1m=00m0, (6)

hence

ϕ(t)={1k=1nrk(t)=2l-n0k=1nrk(t)2l-n.

Therefore

μ{t[0,1]:k=1nrk(t)=2l-n} = 01ϕ(t)𝑑t
= 0112π02πeix(-(2l-n)+k=1nrk(t))𝑑x𝑑t
= 12π02πe-ix(2l-n)01eixk=1nrk(t)𝑑t𝑑x
= 12π02πe-ix(2l-n)cosnxdx;

the last equality uses (4) with c1=x,,cn=x. Furthermore, writing

cosnx=2-n(eix+e-ix)=2-nk=0n(nk)eix(2k-n),

we calculate using (6) that

12π02πe-ix(2l-n)cosnxdx = 2-nk=0n(nk)12π01e-ix(2l-n)eix(2k-n)𝑑x
= 2-nk=0n(nk)12π01eix(2k-2l)𝑑x
= 2-nk=0n(nk)δ2k-2l,0
= 2-nk=0n(nk)δk,l
= 2-n(nl),

proving the claim. ∎

We now prove that the expected value of a product of distinct Rademacher functions is equal to the product of their expected values.33 3 Masayoshi Hata, Problems and Solutions in Real Analysis, p. 185, Solution 13.2.

Theorem 4.

If k1,,kn are positive integers and k1<<kn, then

01rk1(t)rkn(t)𝑑t=0.
Proof.

Write J=01rk1(t)rkn(t)𝑑t and define

ϕ(x)=s=2nr(2ks-k1x),x,

which satisfies

ϕ(x+1)=s=2nr(2ks-k1x+2ks-k1)=s=2nr(2ks-k1x)=ϕ(x).

Hence, as ϕ has period 1 and r has period 2,

J = 01rk1(t)ϕ(2k1t)𝑑t
= 01r(2k1t)ϕ(2k1t)𝑑t
= 12k102k1r(x)ϕ(x)𝑑x
= 12k1j=02k1-1-12j2j+2r(x)ϕ(x)𝑑x
= 12k1j=02k1-1-102r(x)ϕ(x)𝑑x
= 1202r(x)ϕ(x)𝑑x.

But, as ϕ has period 1,

02r(x)ϕ(x)𝑑x=01ϕ(x)𝑑x-12ϕ(x)𝑑x=01ϕ(x)𝑑x-01ϕ(x)𝑑x=0,

hence J=0, proving the claim. ∎

For each n, if f is a function defined on the integers we define

In(f)=01f(k=1nrk(t))𝑑t.
Lemma 5.

For any nN,

In(x2)=n,In(x4)=3n2-2n.
Proof.

Using Theorem 4 we get

In(x2) = 01(k=1nrk(t))2𝑑t
= 01k=1nrk(t)2+jkrj(t)rk(t)dt
= 01k=1nrk(t)2dt
= n.

Using Theorem 4 we get, since rj(t)4=rj(t)2=1 and rj(t)3=rj(t),

In(x4) = 01(k=1nrk(t))4𝑑t
= 01k=1nrk(t)4+(43)j=1nkjrj(t)3rk(t)+(42)j=1nkjrj(t)2rk(t)2
+(42)j=1nj,k,l all distinctrj(t)2rk(t)rl(t)
+j,k,l,m all distinctrj(t)rk(t)rl(t)rm(t)dt
= n+(42)n(n-1).

Our proof of the next identity follows Hata.44 4 Masayoshi Hata, Problems and Solutions in Real Analysis, p. 188, Solution 13.6.

Lemma 6.

For any nN,

In(|x|)=2π01-cosnxx2𝑑x.
Proof.

For n and c1,,cn,

01exp(ik=1nckrk(t))𝑑t=01cos(k=1nckrk(t))𝑑t+i01sin(k=1nckrk(t))𝑑t,

and since (4) tells us that the left-hand side of the above is real, it follows that we can write (4) as

01cos(k=1nckrk(t))𝑑t=k=1ncosck. (7)

Suppose that ξ is a positive real number. Using t=xξ and doing integration by parts,

01-cosxξx2𝑑x = ξ01-costt2𝑑t
= ξ1-cost-t|0+ξ0sintt𝑑t
= ξ0sintt𝑑t
= ξπ2.

It is thus apparent that for any real ξ,

01-cosxξx2𝑑x=|ξ|π2.

For any n, applying the above with ξ=k=1nrk(t) we get

In(|x|) = 2π01|k=1nrk(t)|π2𝑑t
= 2π0101-cos(xk=1nrk(t))x2𝑑x𝑑t
= 2π01x2011-cos(xk=1nrk(t))dtdx
= 2π01x2(1-In(cosx))dx.

Applying (7) with ck=x for each k, this is equal to

2π01x2(1-k=1ncosx)𝑑x=2π01-cosnxx2𝑑x,

completing the proof. ∎

We use the above formula for In(|x|) to obtain an asymptotic formula for In(|x|).55 5 Mark Kac, Statistical Independence in Probability, Analysis and Number Theory, p. 12, Masayoshi Hata, Problems and Solutions in Real Analysis, p. 188, Solution 13.6.

Theorem 7.
In(|x|)2πn.
Proof.

By Lemma 6,

In(|x|)=2π01-cosnxx2𝑑x.

For 0ϵ<1, define ϕϵ:[0,π2) by

ϕϵ(x)=x22(1-ϵ)+logcosx.

We also define

αϵ=arccos1-ϵ,βϵ=αϵ1-cosnxx2𝑑x,

and for σ>0,

Kϵ,σ=0αϵ1-exp(-nx2σ)x2𝑑x.

Let 0<ϵ<1. Until the end of the proof, at which point we take ϵ0, we shall keep ϵ fixed. For 0<x<αϵ we have, using arccos1-ϵϵ,

ϕ0(x)=x22+logcosx<ϵ2+log1-ϵ=ϵ2+12log(1-ϵ)<0,

hence

cosx<exp(-x22).

On the other hand,

ϕϵ(x)=x1-ϵ-tanx,ϕϵ′′(x)=11-ϵ-sec2x,

so ϕϵ(0)=ϕϵ(0)=0 and ϕϵ′′(t)>0 for all 0t<αϵ, giving

ϕϵ(x)>0,

and hence

exp(-x22(1-ϵ))<cosx.

Collecting what we have established so far, for 0<x<αϵ we have

exp(-x22(1-ϵ))<cosx<exp(-x22).

This shows that

Kϵ,2(1-ϵ)=0αϵ1-exp(-nx22(1-ϵ))x2𝑑x0αϵ1-cosnxx2𝑑x,

and therefore

Kϵ,2(1-ϵ)+βϵπ2In(|x|).

On the other hand,

Kϵ,2=0αϵ1-exp(-nx22)x2𝑑x0αϵ1-cosnxx2𝑑x,

so

Kϵ,2+βϵπ2In(|x|).

Now summarizing what we have obtained, we have

Kϵ,2+βϵπ2In(|x|)Kϵ,2(1-ϵ)+βϵ. (8)

For σ>0, doing the change of variable t=nσx,

Kϵ,σ=0αϵ1-exp(-nx2σ)x2𝑑x=nσ0nσαϵ1-e-t2t2𝑑t.

As n, the right-hand side of this is asymptotic to

nσ01-e-t2t2𝑑t=nσπ.

Dividing (8) by n and taking the limsup then gives

lim supnπ2In(|x|)nπ2(1-ϵ),

or

lim supnIn(|x|)n2π(1-ϵ);

indeed βϵ depends on n, but βϵ<2αϵ, which does not depend on n. Taking ϵ0 yields

lim supnIn(|x|)n2π.

On the other hand, taking the liminf of (8) divided by n gives

lim infnπ2In(|x|)nπ2,

or

lim infnIn(|x|)n2π.

Combining the limsup and the liminf inequalities proves the claim. ∎

Lemma 8.

For any ξR and nN,

In(eξ|x|)<In(eξx)+In(e-ξx)=2(coshξ)n.

We will use the following theorem to establish an estimate similar to but weaker than the law of the iterated logarithm.66 6 Masayoshi Hata, Problems and Solutions in Real Analysis, p. 189, Solution 13.7.

Theorem 9.

For any ϵ>0, for almost all t[0,1],

n=11n2+ϵexp(2lognn|k=1nrk(t)|)dt<.
Proof.

Define fn:[0,1](0,) by

fn(t)=1n2+ϵexp(2lognn|k=1nrk(t)|).

Applying Lemma 8 with ξ=2lognn,

01fn(t)𝑑t1n2+ϵ2(cosh2lognn)n.

It is not obvious, but we take as given the asymptotic expansion

(cosh2lognn)n=n-13(logn)2+845(logn)3+118(logn)4n+O(n-3/2),

and using this,

1n2+ϵ2(cosh2lognn)n=2n1+ϵ+O((logn)2n2+ϵ)=2n1+ϵ+O(n-2).

Thus

n=101fn(t)𝑑t=n=1(2n1+ϵ+O(n-2))<.

Because each fn is nonnegative, using this with the monotone convergence theorem gives the claim. ∎

Theorem 10.

For almost all t[0,1],

lim supn|k=1nrk(t)|nlogn2.
Proof.

Let ϵ>0. By Theorem 9, for almost all t[0,1] there is some nt such that nnt implies that fn(t)<1, where we are talking about the functions fn defined in the proof of that theorem; certainly the terms of a convergent series are eventually less than 1. That is, for almost all t[0,1] there is some nt such that nnt implies that (taking logarithms),

(-2-ϵ)logn+2lognn|k=1nrk(t)|<0,

and rearranging,

|k=1nrk(t)|nlogn<2+ϵ2=2+ϵ.

For each s, let Es be those t[0,1] such that

lim supn|k=1nrk(t)|nlogn>2+1s.

For each s, taking 0<ϵ<1s we get that almost all t[0,1] do not belong to Es. That is, for each s, the set Es has measure 0. Therefore

E=s=1Es

has measure 0. That is, for almost all t[0,1], for all s we have tEs, i.e.

lim supn|k=1nrk(t)|nlogn2+1s,

and this holding for all s yields

lim supn|k=1nrk(t)|nlogn2,

completing the proof. ∎

3 Hypercubes

Let mn be Lebesgue measure on n, and let Qn=[0,1]n.77 7 Masayoshi Hata, Problems and Solutions in Real Analysis, p. 161, Solution 11.1.

Theorem 11.

If fC([0,1]), then

limnQnf(x1++xnn)𝑑mn(x)=f(12).
Proof.

Define Xn:Qn by

Xn=x1++xnn,xQn.

We have

QnXn𝑑mn(x)=1nk=1n01xk1𝑑xk=1nk=1n12=12,

and we define

Vn = Qn(Xn-12)2𝑑mn(x)
= Qnk=1nxk2n2+jkxjxkn2-Xn+14dmn(x)
= 1n2k=1n01xk2𝑑xk+1n2j=1nkj01xj𝑑xj01xk𝑑xk-12+14
= 1n2k=1n13+1n2j=1nkj14-14
= 13n+n-14n-14
= n-112.

Suppose that cn is a sequence of positive real numbers tending to 0, and define Jn=Jn(c) to be those xQn such that

|Xn(x)-12|cn.

Then

Vn = Qn(Xn-12)2𝑑mn(x)
Jn(Xn-12)2𝑑mn(x)
Jncn2𝑑mn(x)
= cn2mn(Jn),

so

mn(Jn)Vncn2=n-112cn2.

Take cn=n-1/3, giving

mn(Jn)n-1/312.

Let ϵ>0. Because f is continuous, there is some δ>0 such that |t-12|<δ implies that |f(t)-f(12)|<ϵ; furthermore, we take δ such that

fδ6<ϵ.

Set N>δ-3. For nN and xQnJn,

|Xn(x)-12|<cn=n-1/3N-1/3<δ,

and so

|(f(Xn(x))-f(12)|<ϵ.

This gives us

|Qnf(Xn(x))𝑑mn(x)-f(12)| = |Qnf(Xn(x))-f(12)dmn(x)|
Jn|f(Xn(x))-f(12)|𝑑mn(x)
+QnJn|f(Xn(x))-f(12)|𝑑mn(x)
Jn2f𝑑mn(x)+QnJnϵ𝑑mn(x)
2fmn(Jn)+ϵ
2fn-1/312+ϵ
< fδ6+ϵ
< 2ϵ,

which proves the claim. ∎